Insect detection and counting from entomological collections using deep learning methods

Entomological collections are invaluable repositories of biodiversity records, crucial for
understanding the temporal and spatial distribution of insects, especially important given
current concerns about the decline of insect populations. Despite ongoing digitization
efforts in a lot of natural museums, a significant challenge remains in linking metadata to
individual insect specimens stored in collection boxes. The automated detection of an
insect specimen from a collection box can be a difficult task owing to the remarkable
morphological diversity inherent to these organisms. The advent of convolutional neural
networks (CNNs) have greatly propelled the field of computer vision, especially in object
detection. In this research, deep learning approaches provide a simple basis for carrying
out the task of insect detection and counting from high-resolution pictures of
entomological collection. YOLOv8 and Faster R-CNN algorithms were selected to detect
and count insect from Lepidoptera and Coleoptera orders by setting-up trained models
over more than 80 insect families from Africa. Then, more than 7,900 pictures were
confronted to pre-trained datasets in order to detect and isolate each insect specimen
and, automatically count the insect number per boxes. A comparisons of both algorithms
is discussed in term of precision and computing resources. Automated detection of
insects in entomological collection pictures could be the first step for their taxonomical
classification. In conclusion, the implementation of deep learning algorithms represents a
significant step forward in the digitization and analysis of entomological collections,
offering promising avenues for enhanced biodiversity research and conservation efforts